Examples

Choose the workflow by its input and output contract. The Supported Models page maps each model family to these contracts.

Text Output

• Text generation: one shared workflow for dense LLM, MoE, VLM, and Omni text output; Phi-4 Multimodal adds the model-specific vision-LoRA merge

• Speculative Decoding: MTP, EAGLE3, DFlash, and DSpark acceleration

• ASR: audio to transcript text with Qwen3-ASR

• Nemotron-3.5-ASR: audio to transcript text with a model-specific RNN-T runtime

Speech Output

• Qwen3-TTS: text, style, language, or reference speech to 24 kHz speech

• Qwen3-Omni: text, images, or audio to text and optional 24 kHz speech

Action Output

• Vision-Language-Action: Alpamayo trajectory generation and Cosmos3 policy action generation

Serving and Evaluation

• Experimental Python API and Server: Python and OpenAI-compatible interfaces

• NeMo Evaluator: evaluate a local OpenAI-compatible server